Hosts: Lenar Kess, Damra Vol. Two researchers say the encrypted reasoning traces frontier APIs hand back to you can be decoded, carried between sessions, and replayed into other models — which turns a cost decision about stateless serving into a question about what your logs have been holding all along. The rest of the day sits underneath that: users reverse-engineering a serving change from output quality, and three separate talks arguing agents need budgets rather than permissions.Ilia Shumailov and Alexander Panfilov on Machine Learning Street Talk — decoded traces carry passwords and API keys past sanitized transcripts, and fabricated traces injected into a window get adopted as the model's own prior thinking. Every request-side filter is watching the wrong surface.Claude Code users report effort levels dropping and the Hacker News thread — Anthropic's Thariq confirms a live test that remaps the numerical effort value. The number you type didn't change; its meaning did.Sachin Malhotra on agent budgets — an agent deleted 200 workloads in 90 seconds, and the design that followed classifies calls by whether anybody finds out rather than by what they touch.Token Ops from Microsoft — per-run ledgers with halt and steer policies, including a retriever cut from 20 chunks to 5 mid-run so the answer arrives thinner rather than not at all.DigitalOcean's inference router — 90% correctness against 95% for Opus alone, at fourteen cents a session versus forty-four.Qwen 3.8 on a single RTX 5090 — 77 tokens a second short-context and 64.7 at 128 thousand tokens, which says the key-value cache holds up as the window fills.Vjeux on agent swarms and Rémi Louf's event-driven runtime — one agent writes a wrong fact and the others treat it as ground truth. A content-addressed prompt graph lets you diff two runs and find where the claim entered.Sebastian Fox on clinical note evaluation — roughly 1 note in 20 carries an error serious enough to harm a patient, and the judge model layered on top passes the dangerous ones because there's no oracle for what never reached the page.The AI Daily Brief on data-center politics and Sanders on Mar-a-Lago — the reporting says local opposition is about agency over local decisions rather than power bills, which is a different problem than a messaging problem.Armin Ronacher's Fast and Hard Code and a laid-off designer's replacement for three Adobe apps — the cost that dropped is the ten-thousandth line, not the first.
